WSJT-X/lib/filbig.f90
Joe Taylor 175f1913b4 FFTW wisdom is now built into jt9[.exe].
New optional argument to jt9: -w patience
Default is patience = 1

Example timing measurements for 130610_2343.wav:

patience  plan   execute
          (s)      (s)
-----------------------------------------------
   0      0.01     1.25  FFTW_ESTIMATE
   1      0.69     1.25  FFTW_ESTIMATE_PATIENT
   2     16.97     1.15  FFTW_MEASURE
   3    390.88     1.15  FFTW_PATIENT

Conclusions, consistent with expectation based on past experience
with similar FFTs:
  - First decode (in each mode) with patience = 2 is slow.
  - Speed advantage of patience = 2 is small but measurable.
  - No measurable advantage in using patience > 2.

Present mainwindow.cpp has "-w 1" hard-wired.


git-svn-id: svn+ssh://svn.code.sf.net/p/wsjt/wsjt/branches/wsjtx@4610 ab8295b8-cf94-4d9e-aec4-7959e3be5d79
2014-11-20 18:48:53 +00:00

134 lines
3.8 KiB
Fortran

subroutine filbig(dd,npts,f0,newdat,c4a,n4,sq0)
! Filter and downsample the real data in array dd(npts), sampled at 12000 Hz.
! Output is complex, sampled at 1378.125 Hz.
parameter (NSZ=3413)
parameter (NFFT1=672000,NFFT2=77175)
parameter (NZ2=1000)
real*4 dd(npts) !Input data
real*4 rca(NFFT1)
complex ca(NFFT1/2+1) !FFT of input
complex c4a(NFFT2) !Output data
real*4 s(NZ2)
real*8 df
real halfpulse(8) !Impulse response of filter (one sided)
complex cfilt(NFFT2) !Filter (complex; imag = 0)
real rfilt(NFFT2) !Filter (real)
integer*8 plan1,plan2,plan3
logical first
include 'fftw3.f90'
equivalence (rfilt,cfilt),(rca,ca)
data first/.true./
data halfpulse/114.97547150,36.57879257,-20.93789101, &
5.89886379,1.59355187,-2.49138308,0.60910773,-0.04248129/
common/refspec/dfref,ref(NSZ)
common/patience/npatience
save
if(npts.lt.0) go to 900 !Clean up at end of program
if(first) then
nflags=FFTW_ESTIMATE
if(npatience.eq.1) nflags=FFTW_ESTIMATE_PATIENT
if(npatience.eq.2) nflags=FFTW_MEASURE
if(npatience.eq.3) nflags=FFTW_PATIENT
if(npatience.eq.4) nflags=FFTW_EXHAUSTIVE
! Plan the FFTs just once
call timer('FFTplans ',0)
call sfftw_plan_dft_r2c_1d(plan1,nfft1,rca,rca,nflags)
call sfftw_plan_dft_1d(plan2,nfft2,c4a,c4a,FFTW_FORWARD,nflags)
call sfftw_plan_dft_1d(plan3,nfft2,cfilt,cfilt,FFTW_BACKWARD,nflags)
call timer('FFTplans ',1)
! Convert impulse response to filter function
do i=1,nfft2
cfilt(i)=0.
enddo
fac=0.00625/nfft1
cfilt(1)=fac*halfpulse(1)
do i=2,8
cfilt(i)=fac*halfpulse(i)
cfilt(nfft2+2-i)=fac*halfpulse(i)
enddo
call timer('FFTfilt ',0)
call sfftw_execute(plan3)
call timer('FFTfilt ',1)
base=cfilt(nfft2/2+1)
do i=1,nfft2
rfilt(i)=real(cfilt(i))-base
enddo
df=12000.d0/nfft1
first=.false.
endif
! When new data comes along, we need to compute a new "big FFT"
! If we just have a new f0, continue with the existing data in ca.
if(newdat.ne.0) then
nz=min(npts,nfft1)
rca(1:nz)=dd(1:nz)
rca(nz+1:)=0.
call timer('FFTbig ',0)
call sfftw_execute(plan1)
call timer('FFTbig ',1)
do i=1,NFFT1/2 !Flatten the spectrum
j=nint(i*df/dfref)
if(j.lt.1) j=1
if(j.gt.NSZ) j=NSZ
fac=sqrt(min(30.0,1.0/ref(j)))
ca(i)=conjg(fac * ca(i))
enddo
endif
! NB: f0 is the frequency at which we want our filter centered.
! i0 is the bin number in ca closest to f0.
i0=nint(f0/df) + 1
nh=nfft2/2
do i=1,nh !Copy data into c4a and apply
j=i0+i-1 !the filter function
if(j.ge.1 .and. j.le.nfft1/2+1) then
c4a(i)=rfilt(i)*ca(j)
else
c4a(i)=0.
endif
enddo
do i=nh+1,nfft2
j=i0+i-1-nfft2
! if(j.lt.1) j=j+nfft1 !nfft1 was nfft2
if(j.ge.1) then
c4a(i)=rfilt(i)*ca(j)
else
c4a(i)=rfilt(i)*conjg(ca(2-j))
endif
enddo
nadd=nfft2/NZ2
i=0
do j=1,NZ2
s(j)=0.
do n=1,nadd
i=i+1
s(j)=s(j) + real(c4a(i))**2 + aimag(c4a(i))**2
enddo
enddo
call pctile(s,NZ2,30,sq0)
! Do the short reverse transform, to go back to time domain.
call timer('FFTsmall',0)
call sfftw_execute(plan2)
call timer('FFTsmall',1)
n4=min(npts/8,nfft2)
return
900 call sfftw_destroy_plan(plan1)
call sfftw_destroy_plan(plan2)
call sfftw_destroy_plan(plan3)
return
end subroutine filbig